Modeling Contextual Effects in Developmental Research: Linking Theory and Method in the Study of Social Development

Introduction

From early childhood onward, the social context shapes the development of psychological and behavioral processes. Culture, peer groups, and close relationships, for instance, produce dependencies in social cognitive, affective, and behavioral processes. For example, the development of temperament varies between cultures (see e.g., Chen, Wang, & DeSouza, 2006), experience of peer victimization depends on the ethnic context in school (Graham, 2006), and the level of adolescents' delinquency matches to that of their friends (e.g., Snijders & Baerveldt, 2003). Such examples reinforce the idea that social psychological and behavioral variables vary not only between persons but also between social contexts.

While contextual effects on development are easy to acknowledge at the level of theory, their empirical assessment is more challenging. This lack of linkage is true especially for correlational research that depends largely on the appropriate use of data analytic strategy. For instance, discussions of social selection and socialization in adolescent friendships emerged in the 1970s (Kandel, 1978), but the ability to assess such co-occurring developmental processes with appropriate statistical controls has only recently emerged (see Snijders, 2001). In recent years, the number of accounts acknowledging contextual effects in child and adolescent social development has expanded (see e.g., Laursen, 2005; Little & Card, 2005; Prinstein & Dodge, 2008).

In this section, we present four empirical papers that utilize advanced data analytic techniques designed to assess contextual effects in correlational data. Firstly, Card and Hodges apply the social relational model (SRM; Kenny & La Voie, 1984) to study adolescent aggression and aggressor–victim relationships. Secondly, Masyn and colleagues utilize advances in mixture modeling to develop a new measurement framework for evaluating the nature of psychological constructs; they provide an illustration of this new technique using adolescent delinquency data. Knecht and colleagues as well as Sijtsema and colleagues introduce recent advances in longitudinal social network analysis to examine friendship selection and influence in adolescent delinquency and aggressive behaviors. Together, the papers provide an illustration of some promising data analytic techniques that enable us to focus on social contextual effects in social development (obviously, other techniques are available as well).

In this introductory article, we emphasize the connectedness of developmental theory and data analysis, and aim to aid communication between quantitatively oriented scholars and those interested in applying these techniques. We begin by examining the challenge of assessing social contextual effects in correlational research and proceed with a brief overview of social relational, mixture, and social network modeling.

Context-specificity in Social Development: The Challenge of Data Analysis

Although this section emphasizes the analysis side of the equation, social contextual views on development are driven by theories of human development. For instance, ecological theory posits the effects of multiple contextual factors on human development (Bronfenbrenner, 1979), dyadic theories emphasize the special nature of social interaction in close relationships (see Reis, Collins, & Berscheid, 2000), and perspectives on social norms call attention to groups that share and reinforce psychological and behavioral processes (e.g., Chang, 2004). The interactive nature of person–environment relations is further reflected in the transactional person–environment view of behavioral development (Bandura, 1978) as well as in social information processing models where reactions from others are considered to influence behaviors via their effects on underlying cognitive and affective processes (see Crick & Dodge, 1994; Lemerise & Arsenio, 2000). Moreover, the social selection and influence perspective addresses the active role of individuals' characteristics in the selection of social relationships, which may also further influence characteristics of the interacting individuals (for a timely overview of social influence, see Dishion, Piehler, & Myers, 2008).

Inherent in these theoretical positions is the idea that social development does not take place in a vacuum, but in connection with social others. In spite of the ample theoretical arguments for such social contextual influences, we still too often only collect and analyze social developmental data at the individual level. While experimental social psychology utilizes laboratory manipulations like priming to induce reactions that are inherently context-specific, developmental research uses mostly correlational designs where context-specificity is often an indirect rather than a direct product of measurement. Data from surveying individual participants, for instance, often contain between-subject dependencies in the assessed data because of the shared contexts of data collection like schools, classrooms, families, or friendship relations. In such nested data structures, individuals are likely to resemble each other when they share a similar context. In other words, participants' scores in psychological and behavioral variables often share significant covariation with those of others embedded in the same social context.

Traditional data analysis, however, is predicated on the assumption of statistical independence of observations. When observations share some dependencies, the parameter estimates from traditional procedures are biased (i.e., increasing either Type I or Type II error rates; see Laursen, 2005). Unfortunately, much research on social development overlooks rather than addresses the between-subject dependencies in the data, despite the fact that such effects in the data may have particular theoretical or empirical relevance in the field.

When explicitly acknowledged, such contextual effects in correlational data can either be examined for its theoretical importance or simply controlled for without a specific focus on them. Hierarchical data structures, such as students nested within classrooms, for instance, can be analyzed by controlling for the clustering of cases within the higher-order data units. For example, maximum likelihood estimation of hierarchically ordered survey data can be used to derive unbiased parameter estimates and adjust standard errors to account for the clustering effects (see Muthén & Muthén, 1998–2007). Such clustering effects indicate that the assessed variables vary not only between subjects, but also between the higher-order data units (i.e., the contexts in which individuals are embedded). For instance, early adolescent bullying behaviors vary between individuals as well as between classrooms (see Salmivalli & Voeten, 2004). When present but not controlled for, contextual dependencies in the data confound parameter estimation at the individual level. Conventional regression estimates of individual differences, for example, are typically inflated.

Rather than treating contextual effects as statistical nuisance, assessing them directly will provide fruitful, theoretically bound evidence of the social influences on development—largely in connection with available data analytic techniques. For instance, models of social cognition have long acknowledged both individual and contextual effects on cognitive processes (e.g., Dodge & Murphy, 1984). Empirical evidence for such data patterns, however, did not surface until hierarchical data estimation procedures were used to disentangle the variance in cognitive variables in to the between-subject and between-context levels (see Dodge, Laird, Lochman, & Zelli, 2002; Ojanen, Aunola, & Salmivalli, 2007). Another fruitful illustration may be found in the study of school bullying where bully–victim relationships were originally characterized in terms of a power difference and mutual dislike between the bully and the victim (Olweus, 1978). Here, traditional research has focused on individual rather than relationship-specific effects. Recent findings from dyadic modeling of bully-victim relationships, however, indicate that the likelihood of victimization is indeed stronger in the case of mutual rather than unidirectional antipathies (Card & Hodges, 2007). Moreover, the distinct but complementary perceptions of bullies and victims do seem to involve perceived vulnerability from the part of the victim and dominance from the part of the bully (see Veenstra et al., 2007).

In addition to their theoretical relevance, data analytic techniques also have methodological implications. For instance, to assess dyadic processes in social development, participants must be asked to identify significant others (e.g., friends or classmates) and/or specific perceptions related to significant others (e.g., behaviors, or relationship quality). To examine social selection and influence processes with appropriate statistical controls, collecting longitudinal data on relationship networks where these processes may be assessed with increased precision is required (see Burk, Steglich, & Snijders, 2007).

Clearly, a comprehensive assessment of individual and social contextual effects in social development has implications for the information that can be extracted from the data, as well as for data collection and measurement. In the following sections, we provide a brief overview of three data analytic approaches that may be used to increase our understanding of context-specific effects in child and adolescent social development.

Relational, Mixture, and Social Network Modeling: Capturing the Contexts of Development

Contextual effects on development range from relationship dyads to social groups and evolving networks of social relationships. Each perspective provides insights on the active individual–environment relations and the possible mechanisms through which they shape the development of psychological and behavioral processes. These possible mechanisms may be examined via social relational, mixture, and social network modeling.

Social Relational Modeling: Individual and Dyadic Effects on Development

Dyadic data analytic techniques provide an analysis tool to capture relationship-specific effects on social development. Because social psychological and behavioral processes develop inherently in connection to social others, variables traditionally measured at the individual level are likely to be influenced by relationship-specific processes as well. In fact, dyadic effects on development are often unique at the level of the dyad (i.e., the effect of a relationship on development is often more than just the sum of the two individuals).

The SRM (Kenny & La Voie, 1984) was developed to assess dyadic-level information in a group context. Firstly, information about one's perceptions of social others in a meaningful social context is collected. Secondly, the data are organized into matrices consisting of individuals in the context (e.g., students in a classroom) where the participants' perceptions of each other represent the values to be analyzed (for details on the data matrix arrangement, see Card & Hodges, 2009). Thirdly, the variance in the assessed variables is then separated into actor and partner effects for each individual in the matrix, as well as into dyadic-specific effects. Actor effects reflect one's perceptions of others in general (instead of within the dyad), whereas partner effects reflect others' perceptions of the target child. The type of information that can be retrieved from the data depends on whether univariate (single matrix) or multivariate (multiple matrices) SRMs are conducted (for details about SRM covariance matrix calculations, see Card, Little, & Selig, 2008; Kenny, 1994). The benefit of the SRM approach is that contextual effects on development are assessed from multiple perspectives. In fact, whether one's perceptions of others correspond to those held by others may have significant implications for social adjustment (see e.g., Card & Hodges, 2007, 2009).

To date, the SRM technique has mostly been used to study inter-personal perception in adults (for a review, see Kenny, 1994). However, SRM applications have recently surfaced in the developmental field as well (e.g., Card & Hodges, 2007; Ross, Stein, Trabasso, Woody, & Ross, 2005), as have applications of its methodological sibling, the actor–partner independence model (APIM; Kashy & Kenny, 2000). Both techniques provide insights into the unique effects of social relationships (inter-dyadic differences) and inter-individual differences on development. The SRM assesses dyads in the group context, whereas the APIM focuses on mutual actor and partner effects within individual dyads (for a comparison of these techniques, see Little and Card, 2005).

In this issue, Card and Hodges apply the SRM to the study of adolescent aggression and victimization to assess between-person and between-relationship differences in these variables. After collecting data on aggression directed toward others and aggression received from others (two scales of the newly established dyadic aggression and victimization inventory or DAVI), Card and Hodges utilize multivariate SRM to examine the internal consistency of actor, partner, and dyadic effects in aggression and victimization; these three aspects of reliability may be assessed only with the multivariate application of the SRM. As reflected in the findings, adequate reliability of the new instrument, together with validity evidence, which is evaluated with respect to other measures of aggression and victimization, make the DAVI the first empirically established method for assessing aggressor–victim dyads in middle school. The findings of Card and Hodges (2009) underline the relatively strong impact of the aggressor–victim dyads on adolescent social adjustment and have theoretical and methodological implications for future research in the area.

Mixture Modeling: An Empirical and Methodological Tool to Study Social Development

While social relational modeling provides information about dyadic effects on development, mixture modeling can be used to identify participants that share psychological or behavioral characteristics. As in traditional cluster analysis, distinct data profiles are identified based on participants' scores on particular variables (the term ‘mixture modeling’ refers literally to the assessment of mixtures of score distributions). Perhaps the most common term associated with mixture modeling in the developmental field is that of developmental ‘trajectory’. For instance, developmental trajectories of social withdrawn behaviors and associated adjustment variables provide valuable information about the developmental risks and protective factors for shy and withdrawn children (Oh et al., 2008). Another commonly used term in the mixture modeling framework is ‘latent class’, referring to the data profile as a latent variable in cross-sectional or longitudinal data.

Mixture modeling in the social sciences shifted from non-parametric estimation of response sequences (see Rindskopf, 1990) to the use of semi-parametric estimation techniques designed to identify homogeneous clusters of developmental trajectories (see Nagin, 1999). In the latter approach, the trajectories are identified based on mean score distributions via a chi-square test of homogeneity and the optimal number of groups is estimated with the Bayesian information criterion (given that the traditional likelihood ratio test is suitable only for nested model comparisons; see Nagin, 1999). Furthermore, group membership probability may be linked to other relevant variables via repeated measures multivariate analyses of variance for continuous variables and via log-linear analysis for categorical variables (for an example study utilizing this assessment approach, see e.g., Kokko, Pulkkinen, Mesiainen, & Lyyra, 2008).

More recent work in the field utilizes model-based, parametric data estimation where the trajectories are identified via maximum-likelihood estimation of mean and variance information in the data. Furthermore, the probability of group membership for each participant is calculated with respect to all identified groups and the participants are linked to the groups based on their estimated probability of membership (see Muthén, 2004; Muthén & Muthén, 2000). An additional advantage of this approach is that covariates may be directly included to the model to estimate their relations with class membership (rather than conducting post hoc examinations of these effects). In longitudinal data, this approach is generally referred to as growth mixture modeling, as it incorporates features from both growth models and latent class analysis (see Muthén & Muthén, 1998–2007).

Over the past decade, growth mixture models have been increasingly used to identify trajectories, among others, in the development of internalizing and externalizing behaviors, academic achievement, and personality profiles. In this issue, Masyn and colleagues introduce a novel idea of utilizing mixture modeling as a statistical measurement system (dimensional categorical spectrum) for evaluating the nature of psychological constructs. Specifically, given that the mixture modeling technique may be used to assess both continuous and categorical latent variables in the same model (see Muthén & Asparouhov, 2006) and recent developments in the maximum likelihood estimation enable one to model manifest variables of different measurement modalities (e.g., ordinal or multinominal), this statistical tool may be used in evaluating whether psychological constructs are inherently categorical or dimensional. That is, whether we should examine individual differences in the ‘kinds’ (categorical approach) or the ‘degree’ (continuous approach) of psychological attributes (Masyn et al., 2009) may become a question of data estimation rather theoretical debates.

Masyn and colleagues (this issue) provide an illustration of the newly established measurement framework by assessing adolescent delinquency measured by the child behavior checklist (Achenbach, 1991). By estimating both dimensional and categorical data structures, the authors conclude that a dimensional latent structure is best suitable to describe the delinquency items of this well-established measure. The findings represent an intriguing example of using the mixture modeling technique as a methodological (rather than empirical) tool in the study of social development and provide an initial step toward future research in the area.

Social Network Modeling: Insights from Mathematical Sociology

Social relationships are not only dyadic, but also embedded in larger networks of individual relationships. Such networks encompass meaningful contexts in life, such as schools, neighborhoods, or the family context. Social network analysis addresses the larger scale network structures reflecting multitudes of relationships in a comprehensive assessment framework. Over the past decade, interest in network analysis has dramatically increased across the physical and social sciences (for a review, see Borgatti, Mehra, Brass, & Labiance, 2009).

Accurate mathematical estimation and graphical representations of complex sets of relationships is computationally demanding and analytically challenging. Since the 1970s, the gravity in the multidisciplinary field of network analysis has been in mathematical sociology where complex algebras are continuously introduced to improve the estimation of structural network parameters.

Social network analysis has lot to offer for social developmentalists. In cross-sectional data, it may be used to estimate actor, partner, and relationship-specific effects in social development (see e.g., Veenstra et al., 2007). Longitudinal assessment of relationship networks, in turn, may be used to model changes in compositions of relationships based on individual characteristics (social selection effects) and the network relationships may also be used to predict changes in individual characteristics (social influence effects). In fact, the improved precision in the estimation of selection and influence effects has made this analytic technique ideally suited for evaluating these social developmental processes (for details on data estimation, see Snijders & Baerveldt, 2003).

Central concepts in network analysis include: (a) relationship ties; (b) the network structure; and (c) individual and dyadic covariates. Relationship ties, or nominations among the individuals, are modeled as unidirectional (a nominates b), which allows one to estimate the degree of reciprocity in the nominations (b may or may not return the nomination of a), along with multiple other characteristics. This information may be of particular interest for research on close relationships like friendships, where friendship characteristics can be evaluated in a comprehensive assessment framework. Numerous parameters can be estimated to describe the overall network structure, such as reciprocity and the number of given and received nominations. In the social developmental field, these parameters provide valuable information about the characteristics of child and adolescent peer groups, for instance. In addition, individual and dyadic covariates (variables reflecting individual and relationship-specific effects) can be included to estimate their relations with the network structure (in longitudinal data, these covariates reflect social selection and influence effects). Again, multiple parameters describing these relations can be estimated (for an overview of selection and influence parameters, see Wasserman & Faust, 1994).

Social network analysis offers a diverse analytic tool for assessing social relationships in cross-sectional and longitudinal data. In this special section, Knecht and colleagues as well as Sijtsema and colleagues utilize longitudinal social network modeling to evaluate friendship selection and influence processes during early adolescence. Both papers utilize recent advancements in the field, incorporated in the SIENA module of the StOCNET statistical package (Snijders, Steglich, Schweinberger, & Huisman, 2007). These include the possibility to evaluate parameters across individual networks in the data to increase statistical precision (referred to as multilevel or meta-analysis; see Snijders & Baerveldt, 2003) and improved statistical estimation of co-evolving network and behavioral data (see Burk et al., 2007).

Knecht and colleagues (2010) examine friendship selection and influence in adolescent delinquency. In the absence of social influence effects, their findings underline the role of friendship selection in the development of adolescent delinquency, thus supporting the social control theory of delinquency (Hirschi, 1969). Sijtsema and colleagues (this issue) disentangle friendship selection and influence effects in different dimensions of adolescent aggression. Advancing the forms and functions view of aggression (Little, Jones, Henrich, & Hawley, 2003), the authors investigate the role of instrumental and reactive functions as well as overt and relational forms of aggression in adolescent friendship development. The findings indicate differential selection and socialization effects for the different aggression variables, thus, underlining the importance of treating aggression as a heterogeneous construct in adolescent development.

Assessing the ‘Social’ of Social Development: Today and Tomorrow

As reflected in this special section, best-practice data analytic techniques provide fruitful directions for the study of social development. However, the field also faces challenges with a continuous flow of newly developed techniques and emerging software. In fact, some of the techniques used for these papers will already be significantly updated at the time of publication. For instance, the SIENA software (Snijders et al., 2007) used in social network analysis is currently being revised to enable the assessment of the quality of relationship ties (in addition to the quantity). In addition to improved numerical optimization routines, this new feature of SIENA will greatly enhance the utility of this technique in the study of close relationships like friendships (which inherently include varying levels of affiliation; see Nangle, Erdley, Newman, Mason, & Carpenter, 2003).

In the continuously developing field, evaluating and comparing differential data analytic techniques is warranted (see Kindermann & Gest, 2009). For instance, both multilevel and social network modeling may be used to evaluate friendship selection and influence processes (for an example study comparing these approaches, see Burk, Kiuru, Laursen, Nurmi, & Salmela-Aro, 2009). While scrutiny is in order, the current stage of development in the quantitative field enables us to capture the ‘social’ of social development in more detail today than ever before. We hope that this special section provides helpful insights on the uses of some of these techniques and inspires more researchers to directly examine the social in social development.